The Claude Operator Pack is four production workflows for running Claude, or any frontier model with tool access, as an operator: a system that researches, verifies, delegates, and files its work, not one that just answers questions. Each workflow is a self-contained SKILL.md procedure. Drop the folder into a Claude Code or agent skills directory, paste one into a Claude Project's instructions, or adapt it to any agent framework, and it runs the same way every time.

These are not prompts written for a demo. The pack was built and is used daily in a live operation: multi-agent research digs, payout verification across dozens of platforms, and a governed document system designed to survive context loss between sessions.

What you get:

• Research Digest: records-first research that scopes to the gaps in what you already know, decomposes a question into claim-sized passes, and returns a verdict table with evidence, sources, and confidence grades.

• Verification Pass: a claim-grading procedure (Observed, Source-Verified, Blocked) that names the proxy before trusting it, opens the evidence itself, and reports the split honestly instead of a blended "looks good".

• Multi-Agent Coordination: one coordinator, self-contained worker briefs, a preservation list on every brief, and a reconciliation step that treats every worker's "done" as a claim to be checked.

• Document Production: governed documents with one canonical home per fact, dated rows that supersede in place, and staleness treated as a state, so registers stay true across sessions.

• Four full worked examples, one per workflow, complete outputs on real cases rather than excerpts: a platform-vetting dig, a payout claim set graded tier by tier, an eleven-worker coordination with one brief verbatim, and a live register-correction cycle.

• The combined guide, OPERATOR-PACK.md, with all four workflows in one file for plain reading.

Start with the combined guide to learn the system, install whichever SKILL.md matches the work in front of you, and keep the matching worked example beside it for the first few runs. The examples are the standard the outputs are graded against.